Top 5 CIKM 2023 Papers on Recommender Systems, Search & Datasets

International Conference on Information and Knowledge Management (CIKM) 2023, held in Birmingham, accepted 354 papers (24% acceptance). Five of these papers, authored by researchers from Kuaishou and academic institutions, focus on recommender systems, search ranking, and dataset construction.

Paper 01: SHARK – A Lightweight Model Compression Approach for Large‑scale Recommender Systems

Download: https://arxiv.org/abs/2308.09395

Authors: Zhang Beichuan, Sun Chenggen, Tan Jianchao, Cai Xinjun, Zhao Jun, Miao Mengqi, Yin Kang, Song Chengru, Mu Na, Song Yang (all from Kuaishou)

Increasing the size of the embedding layer improves recommendation performance but can exceed terabytes in industrial systems, raising computation and storage costs. SHARK addresses this by (1) using a Taylor‑expansion‑based approximation of Shuffle to prune the number of embedding tables, and (2) applying a novel quantization method that assigns different quantization strategies to each embedding. Experiments on public and industrial datasets show each component outperforms prior methods. Online A/B tests on Kuaishou short‑video, e‑commerce, and advertising models demonstrate up to 70% memory reduction for the embedding layer and a 30% increase in QPS without performance loss, with all compressed models deployed online.

Paper 02: Query‑dominant User Interest Network for Large‑scale Search Ranking

Download: https://arxiv.org/abs/2310.06444

Authors: Guo Tong, Li Xuanping, Yang Haitao, Liang Xiao, Yuan Yong, Hou Jingyou, Ke Bingqing, Zhang Chao, He Junlin, Zhang Shunyu (all from Kuaishou)

User historical behavior is powerful for recommendation, but search behavior is sparse. Existing personalized search methods rely on sparse search signals, limiting interest modeling. The proposed network consists of two cascaded units: a relevance‑retrieval unit that extracts a sub‑sequence of behaviors relevant to the current query, and a hybrid‑attention unit that computes attention scores for item IDs and attributes, automatically fusing item and content embeddings based on user interaction depth. Offline evaluations on public datasets and online A/B tests on Kuaishou search confirm the method’s effectiveness.

Paper 03: Leveraging Watch‑time Feedback for Short‑Video Recommendations – A Causal Labeling Framework

Download: https://arxiv.org/abs/2306.17426

Authors: Zhang Yang (University of Science and Technology of China), Bai Yimeng (USTC), Chang Jiexin (Kuaishou), Zang Xiaoxue (Kuaishou), Lu Song (Kuaishou), Lv Jing (Kuaishou), Feng Fuli (USTC), Niu Yanan (Kuaishou), Song Yang (Kuaishou)

Short‑video recommendation heavily relies on watch‑time as feedback. Existing methods treat watch‑time as a direct label, ignoring its rich semantics and introducing bias. The proposed DML (Debiased Multi‑semantic Label) framework derives quantile‑based labels from the distribution of watch‑time, emphasizing relative order rather than absolute values. A causal‑adjustment‑inspired technique further mitigates label‑level bias. Both online and offline experiments demonstrate that DML effectively captures true user interests and improves engagement.

Paper 04: Learning and Optimization of Implicit Negative Feedback for Industrial Short‑video Recommender System

Download: https://arxiv.org/abs/2308.13249

Authors: Pan Yunzhu (University of Electronic Science and Technology of China), Li Nian (Tsinghua), Gao Chen (Tsinghua), Chang Jiexin (Kuaishou), Niu Yanan (Kuaishou), Song Yang (Kuaishou), Jin Depeng (Tsinghua), Li Yong (Tsinghua)

In short‑video recommendation, skip behavior is a prevalent implicit negative signal that poses two challenges: (1) it reflects nuanced user preferences, making interest extraction difficult, and (2) it affects total watch‑time, influencing multiple business objectives. The solution deploys a feedback‑aware encoding module to capture user preferences with contextual awareness, followed by a multi‑objective prediction module that disentangles different optimization goals. Extensive online testing and analysis confirm the approach’s effectiveness.

Paper 05: KuaiSAR – A Unified Search and Recommendation Dataset

Download: https://arxiv.org/abs/2306.07705

Code: https://kuaisar.github.io/

Authors: Sun Zhongxiang (Renmin University of China), Si Zihua (Renmin University of China), Zang Xiaoxue (Kuaishou), Leng Dewei (Kuaishou), Niu Yanan (Kuaishou), Song Yang (Kuaishou), Zhang Xiao (Renmin University of China), Xu Jun (Renmin University of China)

KuaiSAR is a unified dataset collected from the Kuaishou short‑video app, containing real user behavior logs for both search and recommendation services over 19 days and 25,877 users (over 300 M daily active users). It records every interaction in search or recommendation, including the transition between the two services, and provides rich auxiliary information for users and videos, as well as both positive and negative feedback. Compared with existing datasets, KuaiSAR is the first to capture authentic search‑and‑recommendation behavior, includes the source of search queries (typed vs. recommended), and documents user‑video interactions comprehensively.

The dataset enables research on unified search‑and‑recommendation modeling, cross‑service user behavior analysis, and multi‑objective optimization in large‑scale short‑video platforms.